2019 stock assessment developments

[jysullivan]

Data

• [X] Work with goundfish biologists to determine if we should account for a change in selectivity or catchability when the fishery changed from open access to limited entry or any other points since 1980
• [X] Work with groundfish biologists to obtain the raw mark-recapture data from 2003 and 2004. 2003 was lost, 2004 used PIT tags.
• [X] Use data to estimate variances and effective sample sizes for the various abundance indices and age compositions. Currently the sigmas for the abundance indices and effective sample sizes for the age comps are assumed constant values, which is likely unrealistic.
• [x] Age comps - I used the square root of the total number of fish aged in a year as the effective sample size. Should revisit in the future. LONG-TERM
• [x] Work with groundfish staff to assess caveats in survey and fishery CPUE (these indices do not appear to follow population trends well).
• [x] Assess whether we should estimate a separate q pre-1985 (before shift to Limited Entry).
• [x] Determine if code for defining length bins is equivalent to Dana's.
• [x] Recommendation by Dana to use the se/mu instead of sd/mu to calculate CV for CPUE indices.
• [x] Plot and incorporate ageing error and age-length keys from Dana. Longer term goal to develop Chatham-specific matrices. Fish length comps will need separate age-length key because samples only represent landed catch.
• [x] Change plus group to age 31!

Model

• [x] Biological reference points not estimating, resulting in Hessian not positive definite.
• [x] Recommendation from Dana to use 0.5 as sex ratio instead of survey sex ratio, which includes selectivity and not true population sex ratio. Caveat that this may not be the best way, but it's at least standard practice. When making this change output N matrix sex ratio to see the impact.
• [x] Add ageing error matrix. Use matrix instead of elemental multiplication.
• [x] Because fishing mortality is continuous in the model, add F to survival fractions and SBPR calcs.
• [x] ABC and wastage calculation incorrectly references current year instead of forecast year.
• [x] Time blocks for fishery q (shift to Equal Quota Share in 1994).
• [x] The model is not getting the scale correct
• [x] Fs are tiny and biomass is enormous – moving to a prior on survey q may help resolve this (perhaps the q for the mark-recapture estimates, which one hopes are close to 1)
• [x] The age-2 recruits seem far too stable – that may be a consequence of the weighting or just the enormous last values. The standard errors are very large and that may reflect that the data are uninformative about scale.
• [x] Mark-recap q may be poorly estimated and the main source of information on scale. If so there should be a direct tradeoff between q, M (if estimated), and mean recruitment. The gradient structure will show the parameter correlation, as would posterior correlation in MCMC samples
• [ ] Weighting factors – Andre suspects that the weighting on the rec_devs is too high and on the F_devs too low – one wants the model to estimate annual recruitment and also fit the catches
• [x] Refine the implementation of phasing f25e448
• [x] Refine starting values
• [x] Add discard mortality as a switch ea58a6b
• [x] Make sure fishery age comps relate back to landed portion of the catch (not total)
• [x] Add switch for Dirichlet-multinomial as alternative for comps c68d675
• [x] Add length compositions LONG-TERM
• [x] Develop sex-structured model to deal with the fact that at-age females are larger than males N matrix but still allow the flexibility of a single-sex model - Update 2019/04/17: getting somewhat reasonable output. having a hard time estimating fishery selectivity.
• [ ] Conduct projections to evaluate stock status and future spawning stock size LONG-TERM
• [ ] Develop code to conduct a retrospective analysis LONG-TERM
• [ ] Develop priors and move towards a Bayesian implementation of this assessment
• [x] Change selectivity (or add another option) for a50/k parameterization, vectorize parameters, Created switch see 163e0cc
• [x] Split rec_devs and rinit_devs into separate variables
• [X] Estimate recruitment as a random effect
• [X] Time-varying selectivity
• [X] Develop the TMB code to estimate biological reference points within the model

Sensitivity analysis

These are fixed in the model (maturity is estimated using NSEI data, but selectivity is borrowed (having a really hard time fitting age comps - I think because of the discarding issue).

• [ ] Maturity
• [ ] Selectivity
• [ ] Impact of a triennial

Text and figures

• [x] Incorporate changes to methods for the assessment in the Appendix to the 2018 assessment due end of June.
• [X] Fix typos in the equations (e.g. not “i” on the RHS for the recruitment model) (ongoing probably)
• [X] Do not use E for predicted CPUE because it can be confused with effort
• [X] Include the assumed standard errors on plots that show fitted values to aid in the evaluation of model fit
• [X] Use standardized instead of raw residuals
• [X] Include zeros on the plots to help place the trends in context

[jysullivan]

Questions

@GreatDana

• [x] What assumptions are made for spawning biomass calculation and sex ratio? Currently I have it coded to use the proportion female-at-age from the survey (only applied in the single sex mode). A: The Fed model hasn't been single sex since Sigler was lead. He estimated a total SSB and then assumed 50% was female.
• [x] For sex-structured model with sex-specific weight-at-age and age comps, what assumptions do you make about the sex ratio of initial numbers-at-age and recruitment? A: Assumes 0.5 sex ratio for both initial numbers-at-age and recruitment.
• [x] Do you estimate init_N_devs and rec_devs in the same or separate vectors? I separated mine, but they have the same sigma_r. A: They are estimated as a single vector, with the same sigma_r.
• [x] Would you check the implementation of the ageing error matrix? It's currently causing the comps to not sum to 1 when I include it (at least in debug mode), so I'm not sure it's right. A: Checked, use matrix multiplication instead of element-wise multiplication.
• [x] Advice/help on an age-length key? Or could we use the Federal one as a starting point? Would love to include length comps before presenting this model. A: Shared Federal age-length and ageing error matrices for short-term use until ADFG can develop our own.
• [x] Would you review my BRPs please? A: Reviewed, recommendations added to TO DO list under section on Modeling.
• [x] Are your selectivity curves fixed or estimated annually/periodically? Will you please provide the parameters used/estimated for the 2018 assessment? A: Selectivity parameters estimated annually.
• [x] Do you have any recommendations on units? My model is jumping between numbers to numbers in millions, and mt to kg. A: Yes to using mt and kg.

Andre

• [X] Check likelihood for random effects (negative sign correct?)
• [X] Estimating init_N_devs and rec_devs in the same vector as a random effect (seems like it could skew it)
• [X] Grant recommended against estimate F_devs as random effects because it could lead to underestimation via shrinkage. Not a problem since we don't need a variance parameter estimated?

[jysullivan]

Notes for the Dirichlet-multinomial:

Merrill wrote a nice implementation of the Dirichlet-multinomial likelihood in LIME https://github.com/merrillrudd/LIME/blob/e7be6a4355fe0dea219d657198e7f7d4a225c9a3/src/LIME.cpp#L557

The reference for this is Eqn 10 in Thorson, James T., et al. "Model-based estimates of effective sample size in stock assessment models using the Dirichlet-multinomial distribution." Fisheries Research 192 (2017): 84-93.

Note that Eqn 10 is the likelihood not the negative log likelihood and the big N is a typo

[jysullivan]

Age compositions

Here are our fishery and survey samples sizes for age compositions. Given the n=500 rule (Thompson 2002), we may have sufficient sample sizes to support sex-structure for the fishery, but not for the survey. Given this information, I think we should keep the compositions combined for now and revisit once I have a chance to talk to other authors at Plan Team.

Source	Sex	1997	1998	1999	2000	2001	2002	2003	2004	2005	2006	2007	2008	2009	2010	2011	2012	2013	2014	2015	2016	2017	2018
LL fishery	Female						1267	1029	1544	1247	1076	986	1015	1046	980	919	924	656	777	753	924	830	813
LL fishery	Male						1028	768	771	503	605	510	517	514	548	547	600	654	561	452	502	590	503
LL fishery	Sex combined						2295	1797	2315	1750	1681	1496	1532	1560	1528	1466	1524	1310	1338	1205	1426	1420	1316
LL survey	Female	264	198	156	216	183	304	391	410	512	420	437	340	366	320	365	407	261	289	293	318	351	353
LL survey	Male	273	158	184	193	275	362	399	367	369	319	339	273	253	285	311	319	289	276	204	238	260	228
LL survey	Sex combined	537	356	340	409	458	666	790	777	881	739	776	613	619	605	676	726	550	565	497	556	611	581

[GreatDana]

Length bin structure (double checked 4/19/19)

How the Feds define it: fishlen$mid<-floor(fishlen$LENGTH/2)*2+1 fishlen[fishlen$mid>99,]$mid<-99 fishlen<-fishlen[fishlen$mid>39,] fishlen<-fishlen[fishlen$mid<100,]

Is equivalent to how we define it: fishle%>%nfilter(!c(length < 40)) %>% mutate(length2 = ifelse(length < 41, 41, ifelse(length > 99, 99, length)), length_bin = cut(length2, breaks = seq(39.9, 99.9, 2), labels = paste(seq(41, 99, 2))))

[jysullivan]

Sex ratio to initial N matrix

Based on conversation with Dana, it is more of a standard practice to use 50/50 sex ratio instead of the sex ratio from the survey to initialize N matrix and estimate recruitment.

Below is a comparison of the two methods looking at the 2018 total population numbers-at-age estimates. Definitely think the latter is more realistic.

Method 1: Initialize with sex ratio from the longline survey

Method 2: Initialize using 50/50 sex ratio

[jysullivan]

Notes for tuning age comps:

Description of the process in Muradian et al 2017:

McAllister MK, Ianelli JN (1997) Bayesian stock assessment using catch-age data and the sampling-importance resampling algorithm. Canadian Journal of Fisheries and Aquatic Sciences 54: 284–300.

Stewart IJ, Hamel OS (2014) Bootstrapping of sample sizes for length- or age-composition data used in stock assessments. Canadian Journal of Fisheries and Aquatic Sciences 71: 581–588.